3-D face recognition

Eriksson, Anders (1999-12)

Thesis (MEng) -- Stellenbosch University , 1999.

Thesis

ENGLISH ABSTRACT: In recent years face recognition has been a focus of intensive research but has still not achieved its full potential, mainly due to the limited abilities of existing systems to cope with varying pose and illumination. The most popular techniques to overcome this problem are the use of 3-D models or stereo information as this provides a system with the necessary information about the human face to ensure good recognition performance on faces with largely varying poses. In this thesis we present a novel approach to view-invariant face recognition that utilizes stereo information extracted from calibrated stereo image pairs. The method is invariant of scaling, rotation and variations in illumination. For each of the training image pairs a number of facial feature points are located in both images using Gabor wavelets. From this, along with the camera calibration information, a sparse 3-D mesh of the face can be constructed. This mesh is then stored along with the Gabor wavelet coefficients at each feature point, resulting in a model that contains both the geometric information of the face as well as its texture, described by the wavelet coefficients. The recognition is then conducted by filtering the test image pair with a Gabor filter bank, projecting the stored models feature points onto the image pairs and comparing the Gabor coefficients from the filtered image pairs with the ones stored in the model. The fit is optimised by rotating and translating the 3-D mesh. With this method reliable recognition results were obtained on a database with large variations in pose and illumination.

AFRIKAANSE OPSOMMING: Alhoewel gesigsherkenning die afgelope paar jaar intensief ondersoek is, het dit nog nie sy volle potensiaal bereik nie. Dit kan hoofsaaklik toegeskryf word aan die feit dat huidige stelsels nie aanpasbaar is om verskillende beligting en posisie van die onderwerp te hanteer nie. Die bekendste tegniek om hiervoor te kompenseer is die gebruik van 3-D modelle of stereo inligting. Dit stel die stelsel instaat om akkurate gesigsherkenning te doen op gesigte met groot posisionele variansie. Hierdie werk beskryf 'n nuwe metode om posisie-onafhanklike gesigsherkenning te doen deur gebruik te maak van stereo beeldpare. Die metode is invariant vir skalering, rotasie en veranderinge in beligting. 'n Aantal gesigspatrone word gevind in elke beeldpaar van die oplei-data deur gebruik te maak van Gabor filters. Hierdie patrone en kamera kalibrasie inligting word gebruik om 'n 3-D raamwerk van die gesig te konstrueer. Die gesigmodel wat gebruik word om toetsbeelde te klassifiseer bestaan uit die gesigraamwerk en die Gabor filter koeffisiente by elke patroonpunt. Klassifisering van 'n toetsbeeldpaar word gedoen deur die toetsbeelde te filter met 'n Gabor filterbank. Die gestoorde modelpatroonpunte word dan geprojekteer op die beeldpaar en die Gabor koeffisiente van die gefilterde beelde word dan vergelyk met die koeffisiente wat gestoor is in die model. Die passing word geoptimeer deur rotosie en translasie van die 3-D raamwerk. Die studie het getoon dat hierdie metode akkurate resultate verskaf vir 'n databasis met 'n groot variansie in posisie en beligting.

Please refer to this item in SUNScholar by using the following persistent URL: http://hdl.handle.net/10019.1/51090
This item appears in the following collections: